Apologies that things have been a bit slow on the Avogadro development side. The last few weeks, I've been pushing ahead on a research project: bouquet
Bouquet is designed to quickly find low-energy conformers using a range of different methods (force-fields, ML potentials, GFN1 / GFN2, DFT) using Bayesian optimization and some additional Monte Carlo moves (e.g., collective motion).
At the moment, it's intended to fill a gap between:
- generating a ton of conformers with RDKit and minimizing them with your method of choice (i.e., hoping you'll minimize to the low energy conformer)
- running CREST or GOAT to generate a full ensemble
It's already available to run, but I hope to have a preprint and an Avogadro plugin ready by the end of the month after I finish more benchmarking.
Discuss this on our forum.
Very cool. Maybe I should wait for the paper, but I guess finding an acquisition function with the right balance of exploration and exploitation could be an interesting exercise?
The BOSS tool from Turku / Aalto goes for that approach. A few of their papers discuss pushing the acquistion function to explore more over time.
We split into two potential tasks:
In particular, we want to make sure Task 1 scales well to larger molecules. Right now, it’s hard to be sure you have a good minima for oligomers / polymers (e.g., CREST seems to struggle to find a helix with alanine oligomers).
Our previous Bayesian optimization work struggled with more torsions (e.g., it did great up to about 8 dihedrals) and we’re now tackling 35+ fairly well.
I can’t (yet) guarantee we get the full ensemble - because we wanted to make sure we got the right minima first.
Nice. I’ll have a read of the BOSS papers - looking forward to a bouquet preprint too.